New Construction Starts data is published monthly by Dodge Data and Analytics. Starts data captures a share of the total market or a portion of all construction, on average about 50% to 60% of all construction. Changes in sample size can introduce potential errors in forecast when using starts to predict construction spending.

In any survey, if sample size remains constant, let’s say at 50% of actual output, but survey response increases 5%/year, then that reflects output should increase at 5%/year. However, if survey response increases at 5%/year but sample size is increasing at 3%/year (50%, 53%, 56%, 59%, etc.) then actual output should increase at only 2%/year.

For a survey sample to be used to compare to itself from year to year to predict growth in spending, sample size must remain constant from year to year. If it is not constant, the apparent growth in starts does not all reflect real growth in spending.

It is impossible within a single year to verify if the current market share captured is constant with previous year sample size. The sample period of data is a year of new starts. To find out if the sample size is consistent, the sample must be compared to actual spending from starts from that period. Starts from any given year get spent over a period of the next 2 to 4 years. It takes several years to see the pattern of starts sample size versus actual spending. 

An average spending pattern for nonresidential buildings starts, OR A TYPICAL CASH FLOW CURVE, for any given year is: 20% of the revenue gets spent in the year started, 50% in the next year and 30% in the 3^rd and 4^th year. Multi-billion $ highway projects, manufacturing facilities, power projects and transportation terminals would have much longer duration cash flow curves. In other words, if you desire to predict construction spending in 2019, you need to know what starts were at a minimum in 2017 and 2018, and in many cases back to 2016 or even 2015.

2018 construction starts do not provide enough information to predict 2019 spending.

If starts survey sample size varies from year to year, it’s possible some of the spending growth anticipated from new starts may not represent growth in real volume of future work but could simply represent a change in sample size. Potential significant variations in sample size are seen in the data and may cause errors in the forecast.

Here are some examples. In the following table the line item “starts vs actual cash flow $” uses cash flow curves unique to each type of construction. For instance, in Office and Educational the spending curve is close to the average 20%/50%/30% as described above. That means 2015 starts is compared to a cash flow curve that spreads spending of 2015 starts over the next three years by 20%/50%/30%.

In the Educational data we see it is unusual that Starts and Backlog continued to grow for five years but that same rate of growth was not reflected in actual spending. From 2013 to 2018 new starts increased more than 60% but spending for the period of those starts (97% gets spent between 2014-2020) increased only 30%. That would seem to indicate a very large volume of work is growing in backlog, and spending, at some point, should boom and remain high for an extended period. But the cash flow model is not in agreement.

A possible explanation is the sample survey of new starts has been increasing, so not all the starts growth for five years represents growth in new work. Some of the increase in starts is simply growth in sample size.

As evidence, Educational starts for the period 2012-2015 averaged just less than 50% sample size of actual total spending. In 2016-2018 the average sample size vs spending was over 60%.

Office Spending increased by 20%/year from 2013 to 2016, but in 2017 it turned to a 1% decline. That was unusual and unexpected since 2016 starts and 2017 backlog had both reached 10-year highs. Highly probable is that the sample size of starts increased dramatically in 2016 and the increase in starts was not all growth in real volume but was partially just a change in sample size, therefore the 2017 spending forecast may have been significantly overstated.

For the period 2011-2015 sample size increased from 45% to near 50% of actual total spending. In 2016, sample size jumped 25%! For 2016-2018 the average sample size vs spending was near 60%.

Transportation Terminals and Rail starts reached record high in 2017, both up 120% after a 35% increase in 2016. Starting Backlog increased 22% in 2017 then jumped 95% in 2018. Spending in 2018 is forecast to finish up more than 20%. However, Transportation sample size of new starts may have increased far more than any other market. Does it all represent a real increase in future spending or is this a good example of a change in sample size?

For the period 2011-2015 sample size increased from 25% to 30% of actual total spending. In 2016, sample size jumped to 40% of actual. In 2017 sample size jumped to 70%!

A large portion of the 2017 increase in starts is expected to be a change in sample size. Starts more than doubled from 2015 to 2017. If all that represented an increase in volume, spending would have doubled from 2016 to 2019. We already have actual spending in hand of more than half of 2017 starts and there is no possible outcome that shows the 125% increase in new starts in 2017 will produce an equivalent increase in spending. Most of the actual spending occurs in 2018 and 2019. For those two years, spending will be up 35%.

Office in 2016 posted a 31% increase in starts, mostly due to Hudson Yards and Vanderbilt Tower in NYC. This appears to have increased the annual share of market captured in the starts for 2016. Overall spending in the following years did not increase. Transportation starts in 2017 posted a 121% increase, but almost all of that can be attributed to an increase in market share captured due to $16 billion in starts for LaGuardia, Orlando and LAX airport work. In a year when several multi-billion $ projects start, the starts data share of market increases. This signifies a change in survey size, not an equally sizable increase in future construction spending.

These examples show that starts share of market captured from year to year are not all consistent and therefore starts compared to previous year should not be used to predict spending directly but that starts sample size must be analyzed before using the data to forecast future spending. Construction Analytics models adjusted starts using unique cash flow curves to predict construction spending for the Economic Forecast published here.